A classifier that can now extract text
Decider does not write. Given a document, a ticket, a page or an image, plus a set of closed questions about it, it returns a probability for each answer: yes or no, one option out of as many as 255, or a level on a rubric. Applications use those numbers to route, filter, moderate, rerank and grade, work that teams have been paying generative models to do and then parsing back out of prose.
Version 1.3 adds the one thing a probability cannot give: a location. A Locate question asks the model to find the text itself, a name, an email address, the city in a job posting, and it returns the exact span with its start and end character positions and a confidence score, or an empty list when nothing matches. That turns the model from a judge into a pointer. For personal-information detection and extractive question answering, the span is the deliverable, and a probability alone was never going to be enough.
Perplexity built Locate as a separate adapter rather than merging it into the model, because merging cost accuracy on ordinary decision questions. The cost of that choice is in the documentation: mix Locate questions with decision questions in one request and the work runs in separate forward passes. Locate also accepts text only, no images, caps a request at 128 questions, and was trained on inputs up to 8K tokens and evaluated up to 32K. Longer inputs are accepted. Their span quality is, in the documentation's own framing, not measured.
The change underneath: questions that stay in their lanes
The quieter fix may matter more in production. In v1.2, every part of a request could attend to every other part, and Perplexity states the consequence plainly: an answer to one question could change depending on what else was asked. A classifier whose verdict shifts because a second check was added to the same request is hard to threshold and harder to audit.
In v1.3, each question attends to the shared state and itself, while the shared state attends to all questions. Questions can therefore influence one another indirectly through that state, despite having no direct question-to-question attention. Perplexity's stability test supports the design, on its own numbers: reorder the options, reassign the labels, rename them, and Decider v1.3 changed its semantic answer on 0.27% of questions, where Jev 1.14 changed on 2.69–4.84%. Send the identical input again and Decider changed nothing; Jev changed on 1.88%. Those are Perplexity's tests, run by Perplexity, against the rival its launch post names most often. They are also the right tests to ask for, and any team evaluating either model can rerun them on its own cases.
Read 91.6% by its answer key
The launch headline is 91.6% mean accuracy across 63 query-classification tasks, 6.54 percentage points above the classifiers Perplexity runs in production. The next clause in Perplexity's own sentence is the one to keep: the references for those tasks were independently relabeled by a frontier LLM. Perplexity says a frontier LLM independently relabeled the references; that does not establish human-validated ground truth. The score measures agreement with a large model's relabeling of the tasks, checked by the company selling the result.
The production test makes the same point at larger scale and, to Perplexity's credit, says so openly. One million real requests from the API's first 18 days hold no ground-truth labels, so frontier LLM judges supplied the references, and Decider v1.3 agreed with them 93.98% of the time. Against v1.2's 93.68%, the version gain on the workload closest to real use is 0.30 percentage points. Against Jev 1.14's 93.35%, it is 0.63. A reader who stops at 91.6% will size this release wrong in both directions: the benchmark's answer key is a model's relabeling rather than human gold labels, and the 0.30-point production gain measures agreement with LLM judges, a different quantity from benchmark accuracy. The two numbers should not be read interchangeably.
The Locate comparison needs the same care. An average of 69.0 across 17 out-of-domain extraction tasks is a wide lead over 43.3, 41.5 and 33.6, until the sizes are read alongside: the named rivals are a 2B model, a model of undisclosed size and a 0.5B model, set against a 27B model and its 4B distillation. Perplexity reports the lead on tasks it classifies as out of domain, and Flash, the 4B distillation, also scores 63.6, above every baseline listed. The table still offers no size-matched comparison: it does not show that a 27B extraction model beats extraction models of its own size, because none is in the table.
Price and weight are the harder claims to argue with
Two claims in this launch do not depend on an answer key. The API lists at $0.02 per million input tokens, output free, with Flash at $0.01, as of November 12, 2026. Google's Gemini Decision 1, launched the day before in the same category, lists at $0.10. Per-token prices are not total cost, since thresholds, retries and human review decide what a handled case costs, but a fivefold list-price gap in the same week is a fact a buyer can check in two documentation pages.
The second is the weights. Both models are on Hugging Face under Apache 2.0, the full model at about 49 GiB and Flash at about 8.5 GiB, for teams that need the decision step inside their own network. Both model cards still carry an instruction to authenticate to a private repository; that instruction is out of date. ALFW checked on November 12, 2026: both repositories' metadata reports them neither private nor gated, and their READMEs and complete file lists, weight shards included, download without a token. The weights are, in practice, a public download.
Perplexity's own guidance closes the piece it began. The model card warns that one fitted temperature cannot be right for every task, and tells users who rely on the probability values to calibrate on their own traffic and set per-task thresholds. A decision model earns its place when 0.8 means what it says on your cases, not on a relabeled suite. Decider v1.3 gives teams a cheaper, self-hostable way to run that test, and one genuinely new thing to test it on.
Sources
- [1] Perplexity, "Perplexity Decider v1.3: Training, eval, and use cases" (official launch post, November 12, 2026)Read source
- [2] Perplexity, model card: pplx-decider-v1.3 (Hugging Face; backbone, weights size, Locate, limits, calibration)Read source
- [3] Perplexity, model card: pplx-decider-flash-v1.3 (Hugging Face; Qwen3.5-4B, about 8.5 GiB, distilled)Read source
- [4] Perplexity, Decisions API documentation, Models and pricing table (read November 12, 2026; $0.02 / $0.01 per million input tokens, output free, web grounding $0.001 per search)Read source
- [5] Unite.AI, "Perplexity Releases Decider V1.3 Models with Text-Span Extraction" (November 12, 2026; release-date corroboration)Read source